December 2025: NVIDIA/TensorRT-LLM focused on performance optimization and architectural flexibility. Key work includes fused kernels for QK normalization and RoPE in multi-head attention, and support for two-model MTP routing in GLM4 MOE. These changes deliver higher throughput, lower latency, and more deployment options for large language models. No major defects reported; the work aligns with business goals of improved efficiency and scalable inference.

November 2025 monthly summary for NVIDIA/TensorRT-LLM focusing on GLM-related QA improvements. Key outcomes include alignment of GLM accuracy tests with latest metrics and configurations, ensuring test reliability and safer release decisions, and strengthening model evaluation coverage across deployments.

October 2025 monthly summary for nv-auto-deploy/TensorRT-LLM. Focused on delivering a new, configurable penalty-control feature in the sampling pipeline, enabling finer-grained control over presence and frequency penalties by ignoring a configurable number of prompt tokens. No major bugs fixed this month. The work strengthens model tuning capabilities and supports more predictable output behavior across deployments.